Information As A Source of Distraction

Chapter 5. The effect of repeated irrelevant CMS Messaging on the detection of safety-critical messaging

Introduction

This experiment was one of a series to determine how signing within the right-of-way affects driver behavior. The focus of this experiment was to examine the potential for driving-irrelevant information on a CMS to cause drivers to lose respect for traffic-related messages on CMSs.

The present study had two purposes: (1) to document how driver gaze behavior would be affected by different information types on CMSs and (2) to evaluate whether drivers might be more inclined to ignore critical safety-related messages if frequently exposed to driving-irrelevant information on CMSs.

Methods

The FHWA highway driving simulator was used to simulate CMS messaging in a freeway environment. An eight-lane freeway was simulated (four lanes in each direction). An overhead CMS was located every 0.5 mi (0.8 km) over a distance of 48.5 mi (78.1 km). One group of participants was presented with CMSs that displayed frequently changing faces every mile and travel-time messages interspersed between the faces signs. Another group saw blank overhead signs every mile with travel-time messages interspersed between the blank signs. The 97th sign was the same for all drivers. The three-line message on that sign read “ACCIDENT AHEAD/ALL LANES CLOSED/USE NEXT EXIT.” Two major hypotheses were tested: (1)whether drivers looked more at signs with frequently changing salient color images than at blank or travel-time signs, which might suggest the salient images distract drivers from their primary task—monitoring the road ahead and (2) whether drivers exposed to visually salient non-traffic-related messages on overhead signs would habituate to or lose respect for the overhead signs and thus fail to detect a critical instruction to exit the freeway. The relationship between following distance and gaze behavior was also examined.

The Simulator

The simulator’s screen consisted of a 240-degree portion of a cylinder with a radius of 8.9 ft (2.7m). Directly in front of the driver, the design eye point of the simulator was 9.5 ft (3 m) from the screen. The stimuli were projected onto the screen by five Barco™ projectors with resolutions of 2,048 pixels horizontally by 1,536 pixels vertically. Participants sat in a late-model compact sedan as shown in figure 32. The simulator’s motion base was not enabled in this experiment. The car’s instrument panel, steering, brake, and accelerator pedal all functioned in a manner similar to real-world compact cars.

Figure 32. Photo. FHWA highway driving simulator.

The simulated vehicle was equipped with a hidden intercom system to enable communications between the participant and a researcher who ran the experiment from a control room. The researcher in the control room could also view the face video from the eye-tracking system and thereby monitor the participant’s wellbeing.

The Simulation Scenario

A 1.5-mi (1.6-km) section of freeway without overhead signs preceded the first CMS. Each CMS spanned all four lanes and approximated the dimensions of a sign 56 ft (17 m) wide by 8.6 ft (2.5m) high. Because of limitations in the resolution of the simulator’s projectors, all signs in the simulator were oversized so that their legibility distance approximated real-world legibility distances. In this experiment, signs were 1.75 times the size of their real-world equivalent. Thus, the simulated overhead signs were sized to approximate the legibility distance of an overhead sign of approximately 32 ft (10 m) wide by 5 ft (2 m) high. A sign of this size with 0.79-inch (20-mm) pixel pitch would enable a display of 488 pixels horizontally and 76 pixels vertically. Before being made oversize, the simulated letter height of the CMS text was 18 inches (46 cm).

Two scenarios defined the between-groups experimental manipulation. One scenario included human faces on every other CMS. The other scenarios simulated no message (blank) on every other CMS. In both scenarios, travel times to a hypothetical destination were displayed between the signs that defined the experimental manipulation. A typical travel-time message is shown in figure 33.

Figure 33. Photo. Travel time to McLean was displayed once per mile.

Thirty-four faces were displayed in sequence on each of the faces signs. Each face was displayed for 3s, and the entire series of faces repeated throughout the experiment. It has been shown that human faces attract and hold attention as few other stimuli can.^(51,52) The face stimuli were captured from two sources: non-copyrighted celebrity photographs from the Internet and selected faces from the International Affective Picture System.⁽⁵³⁾ On the signs that displayed faces, the backgrounds also varied. Thus, on the approach to any faces sign, the participant might be exposed to four or more faces. The location of the faces on the signs varied, either left, center, or right such that faces photographed from the left side were displayed on the right side of the display, faces photographed straight on were centered on the display, and faces photographed on the right side were displayed on the left side of the display. A representative faces sign is depicted in figure 34.

Figure 34. Photo. Faces serving as salient but driving-irrelevant information.

The 97th sign, with the instruction to exit the freeway, is depicted in figure 35. To make the driving task more realistic and visually demanding, vehicle traffic was simulated. The VISSIM traffic model was used to generate the behavior of vehicles in the traffic stream. Because the random number seed for the traffic model was always the same, all participants were immersed in the same traffic stream. However, because participants controlled their own speed, acceleration, and lane choice, participants could experience different traffic conditions in their immediate surroundings.

Figure 35. Photo. Accident ahead message.

At the beginning of test sessions, 5,000 vehicles/h (1,250 vehicles/h per lane) were generated for6.7 min. Participants were instructed to begin driving 2.5 min into this period of traffic generation. Thereafter, approximately every 3 mi (4 km), 500 vehicles/h would exit the freeway at off ramps, and 500 vehicles/h would merge into the traffic stream from on ramps. One elderly participant was reluctant to drive at the posted speed limit of 65 mi/h (105 km/h) and eventually was passed by the entire traffic stream. That participant was replaced. There were off and on ramps every 1.5 mi (2 km), although the traffic model populated only half of them with traffic.

Participants were instructed to maintain 65 mi/h and to drive in the second lane from the right, except when they wanted to pass. This instruction resulted in most participants staying within the initially generated traffic flow throughout the experiment. The full instructions were as follows:

• You will be driving on a freeway that has four lanes in the direction you are traveling. There will be other vehicles on the road, although traffic should not be heavy. You should obey the law and drive as you normally would. The posted speed limit is 65 miles per hour. Please drive in the second lane from the right except when you want to pass slower moving vehicles. Your destination is McLean. This is a relatively long trip and may take an hour. We will be recording how you drive and where you look. Please drive as you normally would. However, if you arrive at your destination (McLean) within 3 minutes of the time you would arrive if driving at the posted speed limit, you will receive a $10 bonus. Do you have any questions?

For most of the 51 mi (82 km) of simulated freeway, the speed of other vehicles followed the cumulative probability distribution shown in table 17. However, congestion was simulated beginning 2,081 ft (634 m) upstream of sign 6 and again every 15 signs (upstream of signs 21,36, 51, 66, and 81). Upon entering congested areas, each simulated vehicle decelerated at a desired rate of 6.6 ft/s (2.0 m/s). After decelerating in these areas, all other vehicles traveled between 30 and 45 mi/h (72 and 48 km/h) according to the desired cumulative probability distribution shown in table 18. Each congestion area continued for 1,200 ft (366 m) and was followed by a 300-ft (91-m) zone in which vehicles accelerated back toward their normal desired speed. Note that the desired speed or acceleration might not have been realized where another leading vehicle provided impedance. In addition, simulated traffic changed lanes to achieve its desired speed or to prepare to exit. The congested areas were included to challenge participants’ attention and reduce boredom on the drive, which lasted approximately 45 min.

Table 17. Desired speed cumulative probability distribution of the simulated traffic stream.

Table 18. Desired speed cumulative probability distribution at congestion locations.

A short practice session preceded the test session. The original purpose of the practice session was to enable the participants to become accustomed to the handling characteristics of the simulated vehicle. However, pilot testing had shown that some participants thought they were supposed to stay on the freeway, regardless of CMS warnings. Therefore, the training session was modified to ensure that participants knew that it was expected that they should follow instructions on the CMSs. The modified training included a minimum of two CMSs that instructed participants to take the next exit. The practice session instructions were as follows:

• You will be driving on a simulated freeway that has no other traffic. The purpose of this five minute drive is to allow you to familiarize yourself with the steering, acceleration and braking of the simulated vehicle. At the start, you will be stopped in the second lane from the right. When we are ready, I will ask you to accelerate slowly to 25 mile per hour and drive within a lane. You will then be asked to brake to a full stop. After that, you will be asked to accelerate to 65 miles per hour and continue driving in the second lane from the right. There will be a changeable message sign with a traffic alert about a bridge out on your route. Follow the detour instructions on that sign.

In fact, once on the detour, another CMS was encountered that directed participants back onto the original route. In the practice session, if a participant failed to follow a detour instruction, the researcher would urge the participant to follow the directions on the next CMS that he or she encountered. In this way, all participants saw and followed the instructions of at least two detour messages on CMSs before beginning the main test session. Those who initially failed to follow the CMS directions were presented with three or more dynamic messages with detour instructions. The purpose of this practice was to maximize the probability that participants who read the final message sign in the test session (about 48 min later) would feel obligated to follow its instruction.

Eye-Tracking System

The simulator was equipped with a four-camera dashboard-mounted eye-tracking system that sampled at 120 Hz.⁽⁴⁹⁾ The system tracked horizontal gaze direction from approximately the location of the right outside mirror to the left outside mirror and vertical gaze direction from the instrument panel to the top of the windshield. Gaze direction accuracy varied by participant. The mean accuracy of gaze position across participants was 1.6 degrees (radius) with a 0.7‑degree standard deviation. The eye-tracking data (e.g., gaze direction of each eye, head position, etc.) were merged with data from the simulator (e.g., vehicle speed, lane position, and steering wheel position) and the current forward view of the simulation visual scene (approximately 60 degrees horizontal by 40 degrees vertical). In addition, a separate dataset was recorded by the driving simulator of the distance between the front bumper of the participant’s vehicle and the nearest simulated vehicle in the participant vehicle’s forward path.

To quantify when and for how long participants looked at each CMS, a researcher used analysis software to indicate an ROI on individual frames of the recorded video image. An example of an ROI is shown in figure 36 (the halo around the sign). ROIs were created for the first 11 CMSs (signs 0–10), and signs 17–25, 32–40, 47–55, 62–70, 77–85, and 94–96. For each of these CMSs, glances at the ROIs were recorded for the last 10 s before the CMS began to pass out of the driver’s view. This resulted in sampling an equal number of travel-time and experimentally manipulated signs at intervals throughout the session so that trends over time in glance behavior could be observed. Note that none of the zones coded with ROIs coincided with the zones that simulated congested traffic.

Figure 36. Photo. ROI indicated on CMS.

Participants

A total of 32 participants—16 males and 16 females—completed the study. All were licensed drivers from the Washington, DC, metropolitan area. The mean age of participants was 47 years (range 20–85). Twenty-four participants provided interpretable eye-tracking data. Otherwise useable data were obtained from eight participants for whom eye tracking was unsuccessful. The mean age of the 24 participants with good eye-tracking data was 45 years (range 20–79 years) and 11were male. Only one participant reported a mild simulator sickness symptom (headache), and no participant dropped out as a result of simulator sickness.

Results

Throughout this chapter, error bars in the charts and graphs represent 95-percent confidence limits around the means.

Response to the Incident-Related Detour Message

Of the 24 drivers for whom eye-tracking data were also available, 7 failed to respond to the message on the 97th sign by exiting the freeway. These results are shown in table 19. The difference in exit-taking behavior between blank and faces groups was not statistically significant by Fisher’s Exact Test. Because there is an apparent trend, even if non-significant, for more drivers to fail to exit in the faces group, a second test was done that included all drivers who completed the study, regardless of the quality of their eye-tracking data. The data for this test are shown in table 20. With the additional participants included, there was no difference between the group presented with faces and that presented with blank signs (apparent or otherwise).

Table 19. Response to warning to take next exit by drivers for whom eye-tracking data were available.

Table 20. Response to warning to take next exit by all drivers who completed the drive.

The eye-tracking evidence for looking at the CMS is summarized in the following subsections. However, most of the participants were also asked during the post-experiment debriefing if they had read the message on the 97th sign. All participants who took the exit said that they had read the message—which seems reasonable given that no participant had taken any of the preceding 48 exits and the 49th exit differed only in that it was preceded by the accident ahead message. Of the nine drivers who failed to exit, seven were asked if they had read the message, and all but one of those also claimed to have read it. The one driver who claimed not to have noticed the warning was in the blank sign group.

Gaze Behavior

Gaze location was measured for the 10 s prior to reaching the point where the sign passed out of the driver’s field of view. The areas defined by the 10 s approach are hereafter referred to as “data collection zones.”

There are numerous issues to be considered in the analysis of eye-tracking data. The mean accuracy of the gaze location in this study was 1.6 degrees. Because foveal vision is limited to about 2 degrees of visual angle, examination of fine detail requires shifting the gaze to within 2degrees of the details. At an actual or simulated distance of 1,000 ft (305 m), 2 degrees of visual angle includes an area of about 35 ft (11 m) in diameter. On a flat, level road, a driver whose gaze is centered on a vehicle 1,000 ft (305 m) ahead would also include three travel lanes and any sign within 17 ft (5 m) of the center of gaze within foveal vision. The lower edge of the simulated signs was 17 ft (5 m) above the travel lanes. Traveling at 65 mi/h (105 km/h), as participants were instructed to do in this experiment, 10s of travel time would traverse 950 ft (290 m) of roadway.

Because the eye tracker sampled at 120 Hz, there was a new estimate of the center of gaze every 0.008s. Despite the limits in accuracy, precision, and the size of the foveal area subtended at long distances, it can be expected that over time, the average gaze position will fall on the object of visual regard. Therefore the problem is to determine which 120-Hz hits on a target should
be counted and which disregarded as noise or error. The analyses that follow employed threedifferent methods for assessing when participants were looking at a CMS. These methods are referred to as glance, look, and fixation. They were defined as follows:

• Glance. Any accumulation of 12 (12 × 0.008 s = 100 ms) or more hits on an ROI, where a hit is a single 120-Hz center of gaze estimate that falls on the sign’s ROI. This is the least restrictive definition of a gaze at a sign; it maximizes both the probability of correctly detecting a gaze as well as the probability of incorrectly identifying a gaze at something else as a gaze at the sign. For any ROI, this measure was binary; either a glance was recorded or it was not.
  
  
• Look. Any accumulation of 7 or more hits on an ROI within a series of 12 frames (100ms). A look began when this criterion was first met and terminated when the number of hits within the preceding 12 frames dropped below 7. This criterion is more conservative than that for a glance but not as restrictive as the definition of a fixation. Also, unlike a fixation, a single look may include more than one gaze at different areas of a sign.
  
  
• Fixation. Seven consecutive gaze vector hits (60 ms) within a fixation radius of 4percent of the vertical image height (i.e., 15 pixels on the 372-pixel image) and centered on the first of the 7 gaze positions that designated the start of a fixation. The fixation continued until there were six consecutive hits (50 ms) outside the fixation radius. For a simulated object 500 ft (152 m) ahead, the fixation radius subtended a visual angle of about 2degrees. A fixation on an ROI was recorded if the center of the fixation was on the ROI. In practice, the fixation criterion is quite conservative in identifying a gaze at an ROI (relatively low probability of correct identification and relatively high probability of a miss) in part because of the restriction that the gaze should remain within a small area and therefore could not capture pursuit movements.

The convergence of the three ROI gaze criteria on the same conclusion should increase confidence in the research findings.

As indicated in the Eye Tracking section, gaze data were analyzed for a subset of 59 signs. Sixdata collection zones included the final portions of the simulated congestion. Given the proximity of these six zones to simulated congestion, it might be expected that time headways would be shorter there than in the remaining free-flow data collection zones. Because the experimental design includes repeated measurements and the time headways were not normally distributed, a GEE model was used to test for differences in headway as a function of data collection zone type. In this model, a gamma response distribution was assumed, and an identity link function was used. A gamma distribution was used because headways can never be zero or negative. For a discussion of GEE, distribution choices, and link functions, see Stokes, Davis, and Koch.⁽⁵⁴⁾ Mean headway in the 52 free-flow data collection zones, not including the zone with the accident ahead sign, was 335 ft (102 m). In the six zones in proximity to congestion, the mean headway was 241 ft (73 m). This difference was statistically significant, χ²(1) = 15.98, < 0.001. Subsequently, analyses of glance, looks, and fixations were conducted separately for the 6 congestion zones and the remaining 52 free-flow zones. Where differences in gaze patterns were not significantly different between free-flow and congested zones, the data were combined.

Relationship Between Exit Taking and Gaze Behavior

All three measures of gaze behavior were used to assess whether exiting after passing the accident ahead message was related to whether participants directed their gaze to the message. No measure (i.e., glances, glance duration, number of looks, duration of looks, number of fixations, or duration of fixations) was related to whether or not participants took the exit.
Table 21 shows the distribution of participants who took the next exit as a function of whether a glance was recorded. Although one participant who did not exit said she did not see the accident ahead sign, the glance data suggested otherwise. The participant who did not take the exit and did not have a recorded glance claimed to have seen the message and, in fact, attempted to exit but decided that traffic in the right lane prevented him from exiting.

Table 22 summarizes the various gaze metrics related to the critical CMS message. Note that the expected mean durations in table 22 include durations of 0.00001 s for participants who did not glance at the critical sign.

Table 21. Relationship between glance to the accident ahead message and taking exit.

Table 22. Mean counts and mean durations for various gaze metrics to the accident ahead message.

Glance Results

The entire experimental drive lasted about 45 min. With sessions of this length, there was some concern that glance behavior might vary over time because of fatigue or boredom or because the perceived value of the sign content changed over time. To test these hypotheses, the proportion of signs glanced at in each of the seven groupings of zones described in the eye-tracking system section was examined. GEE models were used because the glance probability data contained repeated measurements and did not appear normally distributed. The order analyses assumed a binomial probability distribution with a logit (log odds) link function. The explanatory variables in the model were order (1, 2, 3, 4, 5, 6, or 7) and order squared (to test for linear and curvilinear order effects). The sign displaying the detour warning was excluded from the analyses. Results indicated that there were no linear or curvilinear trends in probability of glancing at signs as a function of order.

When closely following another vehicle, drivers might have less spare visual capacity for attending to CMSs. Because different individuals adopt different car-following strategies and because traffic conditions could vary along the route, time headway was considered in the analyses of the gaze data. Time headways were considered short if less than 1.5s and long if greater than that duration. Over all data collection zones and all drivers, 75 percent of headways were greater than or equal to 1.5 s.

Each data collection zone was classified as containing a glance to the CMS in that zone
(glance = 1) or not. GEE models were used to test the effects of sign content type, time headway, and their interaction on the probability of a glance. These analyses assumed a binomial response distribution with a logit link function.

Preliminary analysis showed there was no difference in the probability of a glance between faces and travel-time signs in either the 6 congested zones or the 52 free-flow zones. Nor were there glance probability differences between congested and free-flow zones. Therefore, zones with faces and travel-time signs were combined into a non-blank class, and the congestion classification was not used. The GEE analysis included sign type (blank or non-blank), headway (short or long), and the interaction of sign type and headway. The interaction was not significant. Participants had a higher probability of glancing toward the non-blank signs (P_e = 0.55) than toward the blank signs (expected probability (P_e = 0.21), χ²(1)= 9.11, p = 0.003). Participants had a higher probability of glancing toward a CMS when driving with a long headway (P_e=0.49) than when driving with a short headway (P_e = 0.38), χ² (1) = 4.07, p = 0.044.

Given that a participant glanced at a CMS, the distances at which the glance started and ended were examined. Glance start distance was not significantly affected by sign content, traffic, or time headway. The mean glance start distance was 701 ft (214 m) before the CMS (25thpercentile = 555 ft (169 m), 75th percentile = 901 ft (275 m)).

Expected mean glance end distance was related to sign content, χ²(2) = 31.97, p < 0.001. As can be seen in figure 37, expected mean glance endings were farthest from blank signs and occurred closest with faces signs. Post hoc tests showed all three end-distance means were significantly different from each other.

1 ft = 0.305 m

Figure 37. Chart. Expected mean distance from CMS at the last glance.

Beginning and ending glance distances were based on individual 120-Hz gaze samples at signs that received a glance. The location of the beginning and end of a glance did not provide an indication of the amount of time participants were glancing at the signs. To assess the amount of time participants gazed at the signs, the glance duration was computed using only those signs within a content type that received a glance. Recall that glance duration is the sum of all 120-Hz samples for which a gaze at the sign was recorded, regardless of when in the data collection zone these samples were taken. As a result, glance duration could be the sum of onecontinuous gaze or multiple 1/120-s gazes several seconds apart. A GEE model with gaze duration as the dependent measure and sign content, headway, and the interaction of sign content and headway as predictors were tested. The GEE model assumed a gamma response distribution with an identity link function. Sign content, headway, and the interaction effects were not significant. The estimated mean glance duration was 0.79 s (25th percentile = 0.50 s, 75thpercentile = 1.08 s).

Look Results

Within each data collection zone, the number of looks at the CMS was calculated as a function of headway. The analysis included time headway as a grouping variable, where headways greater than 1.5 s were classified as long, and headways less than that were classified as short.

GEE models were used for hypothesis testing because the data contained repeated measurements and did not follow a Gaussian distribution.

Look Probability

The probability of at least one look at each sign was analyzed as a function of headway, sign content, and their interaction. Preliminary analyses showed no significant difference in the probability of looks at faces and travel-time signs, so in the GEE model that is reported, those two zones were combined, and the sign content variable became a comparison of blank signs with non-blank signs. The GEE model assumed a binomial distribution with logit link function. As can be seen in figure 38, both main effects were significant: participants were more likely to look at least once at non-blank signs than blank signs, χ² (1) = 4.35, p = 0.037, and at either type of sign when headways were long than when headways were short, χ² (1) = 31.31, p < 0.001.

Figure 38. Chart. Expected probability and 95-percent confidence limits of at least one look at a CMS as a function of sign content (blank or not blank) and mean headway.

Number of Looks

The first model examined whether the number of looks at the CMS differed as a function of sign content and headway. This analysis excluded the six data collections zones where congestion was simulated, because the short headways in those zones attenuated the headway effect. As can be seen in figure 39, there was a significant interaction between sign content and headway, χ²(1) = 9.62, p = 0.002. Participants were more likely to take more looks at faces signs when headways were long than when headways were short, and took few looks at blank signs (regardless of headway). Faces and travel-time sign zones were compared in the same manner as blank and faces zones. In that analysis, the interaction of headway with the sign content did not reach statistical significance, χ² (1) = 3.24, p > 0.05. However, the trend was the same as in the faces zones, with more looks at travel-time messages when headway was long than when it wasshort.

Figure 39. Chart. Estimated mean number of looks and 95-percent confidence limits as a function of sign content and time headway.

Look Duration

The duration of looks at the CMS was computed for all looks. (When the participant did not look at a sign, the duration was coded as missing.) The expected mean look durations computed from GEE models that assumed a gamma distribution with identity link function are shown in
figure 40. An unexpected interaction between sign content and headway was found, χ²(1) = 6.34, p = 0.012, in which blank signs received shorter glances with short headways than with long, whereas travel-time and faces signs received longer glances with short headways than with long. All expected mean look durations fell within a narrow range of 160 to 210 ms, so it is not clear that this statistically significant interaction has practical significance, particularly in light of the low probability of looks with short headways.

Figure 40. Chart. Expected mean look duration as a function of sign content and time headway.

Fixation Results

Fixation Probability

For each participant, the probability of at least one fixation on each CMS was calculated, and each data collection zone was classified as having either short or long mean time headway. Whether or not at least one fixation had occurred (fixation = 0 or 1) was the predicted variable in a GEE model that assumed a binomial response distribution with a logit link function. Predictor variables were headway, sign content, and the interaction of headway and sign content. Preliminary analyses indicated that congestion was not a significant factor so the analysis included all data collection zones and congestion was not included in the model. The only significant factor in the model was headway, χ²(1) = 37.94, p < 0.001. With headways greater than 1.5 s, the probability of at least one fixation on the CMS approached 0.3, whereas with a shorter headway, the probability of a fixation was about 0.1.

Number of Fixations

A GEE model with headway, sign content, and their interaction was used to analyze the number of fixations on each sign, given that at least one fixation was recorded. The GEE models assumed a Poisson response distribution with a log link function. No significant effects were obtained. Overall, for signs that received at least one fixation, the mean number of fixations was 2.27 (95-percent confidence limits 1.84 to 2.80 fixations).

Fixation Duration

Within each data collection zone, the duration of each fixation was calculated. Headway, sign content, and their interaction served as predictor variables in a GEE model that assumed a gamma response distribution with an identity link function. Only the effect of sign content was statistically significant, χ²(2) = 11.20, p = 0.004. As can be seen in figure 41, mean fixation durations were longest for blank signs and shortest for faces signs. It should be noted that although the software algorithm used to identify fixations was set to detect fixations as short as 60 ms, the shortest fixation captured was 110 ms.

Figure 41. Chart. Expected mean and 95-percent confidence limits for fixation duration as a function of data collection zone and headway.

Mean fixation durations were short, and fixations away from the forward roadway of this duration were generally considered safe. Furthermore, for most of the approach distance, the overhead signs fell within what would generally be considered the forward roadway (i.e., within 2 degrees of the horizon in the direction of travel). A few long, potentially unsafe fixations on CMSs were observed, but the percentage of fixations longer than 2s was less than 1 percent. Recent analyses by Liang, Lee, and Yekhshatyan suggest that gaze fixations away from the forward roadway greater than 2 s greatly increase the odds of a crash, whereas glances away from the forward roadway, even those between 1.5 and 2 s, may be associated with no increase in crash risk.⁽⁵⁵⁾ Whether those recent analyses apply to CMSs above the roadway is unclear, because CMSs might have fallen within what those previous researchers considered the forward roadway or road center. The longest CMS fixation identified was 2.86 s. Table 23 shows the frequency distribution for fixation durations for the current study.

Table 23. Distribution of fixation durations.

Table 24 shows the number of fixations with durations greater than 2 s as a function of sign content and time headway.

Table 24. Number of fixations longer than 2 s as a function of time headway and sign content type.

Driving Performance Measures

The plan for this study included two measures of driving performance that might be affected by sign content—possibly because attending to or avoiding attending to content might increase driver workload. These measures were steering entropy and speed. Unfortunately, an upgrade to the driving simulator to increase the realism of the steering wheel feel resulted in a loss of the data channel used to measure steering angle and therefore entropy. Speed was examined to detect either decreases or variability in speed to enable greater attention to sign content.

There were no significant effects of mean speed or standard deviation of speed as a function of either sign group (blank versus faces) or sign content (travel time versus other), or the interaction of these variables.

Discussion

Respect for CMS as TCDs

The majority of drivers heeded the accident ahead message that directed them to take the next exit. Only one of the drivers who failed to take the exit claimed to be unaware of the message. Indeed, several of the drivers who did not take the exit were observed trying to change into the exit lane. This suggests that placing non-traffic-related messages on overhead signs will not necessarily lead drivers to ignore these signs. Likewise, the frequently occurring travel-time messages did not show evidence of leading drivers to ignore more safety or operationally critical messages.

No evidence of habituation or loss of respect for CMS messaging was evidenced in this 45 min drive with CMSs recurring every 0.5 mi (0.8 km). The following two caveats to generalization of this finding are warranted: (1) the base case did not include a 45-min drive in which there was no CMS preceding the warning, and (2) these results apply to a single drive, not weeks, months, or years of driving under signs with irrelevant messages.

Overall, 28 percent of participants who were exposed to 96 CMSs failed to exit. Only oneparticipant claimed not to have read the message, and the glance data indicated that that participant had gazed at the message, although this does not imply the she read it.

Although this study cannot address long-term habituation, it did look for changes in glance behavior over the 45-min drive and did not detect evidence of the beginning of such a change. The study should be taken as positive evidence that frequently occurring CMSs that meet a traffic operations and safety need, such as CMSs for active traffic management, should not induce habituation or loss of respect for TCDs.

Gaze Behavior

Three measures of gaze behavior were used because it was uncertain whether any single measure would unambiguously characterize gaze behavior in the presence of CMSs.

The glance and look measures suggested that drivers are more likely to look at non-blank signs, a finding that makes intuitive sense. All three measures showed a strong effect of headway—when headways are short, drivers are less likely to divert their gaze from the roadway to a CMS. Drivers seem to regulate their gaze behavior according to the demands of the driving situation. Even the frequently changing faces signs did not compel drivers to divert their attention away from the driving task.

The look measure suggested that the faces signs were more likely to receive visual attention when the roadway demands for attention were low (i.e., time headway was greater than 1.5 s). Given that change and that faces are generally considered to have high saliency, the look finding was not unexpected. However, the glance and fixation measures did not reinforce the suggestion that changing faces attract visual attention more than static travel-time messages.

The fixation measure is probably the most conservative of the three gaze criteria. That is, it is less likely to incorrectly identify the focus of visual attention and most likely to miss short gazes at moving objects. Nonetheless, it was surprising to find that the longest fixations were to blank signs and the shortest to faces signs. This effect was small and probably not worth interpreting in the absence of additional replications. In addition, the total time that blank signs were looked at across all drivers and all blank signs was miniscule compared with the total time spent with gazing on non-blank signs. In this study, the amount of time gazing at the sky, grass, or trees was not recorded. If it had been, it is probable that gazes at them would have exceeded 2s. Drivers look away from the road ahead for various reasons. This does not imply that they are not attending to the forward roadway. Fixations on the blank signs may have been random fixations to relieve boredom or maintain overall awareness of the environment, and were not necessarily instances of distraction.

Overall, there was little evidence to suggest that CMSs, regardless of the content, are distracting. Participants appear to have attended to the signs primarily when the visual demands of the primary driving task were low. Furthermore, both measures of gaze duration suggest that drivers do not dwell on signs for periods of time that would compromise safety. The next study examines whether the workload imposed by attention to CMSs detracts from detection of safety-critical events in the roadway.

Drivers appear to glance at CMSs when driving demands are low. This is presumably because drivers give priority to attending to safety-critical aspects of the driving task. When the driving demands are high, drivers have little spare capacity to attend to CMS messages. This implies that when a traffic manager posts what is considered a safety-critical message, that message should be tailored to minimize demands on drivers’ attention. The more that attention demands of a message increase, the less likely drivers are to have sufficient spare capacity to process the message. Several FHWA publications are available to provide guidance on how to minimize the attention demands of CMS messages.^(3,5) Conversely, if drivers expect CMS messages to be noncritical or irrelevant, then the probability that safety-critical messages will be ignored may increase. The experiment described in chapter 6 was intended to test this hypothesis.